Evaluating the effect of topic consideration in identifying communities of rating-based social networks

Evaluating the effect of t opic consideration in identifying c ommuniti es of ra ting-based social networks Ali Reih anian Departme nt of In f ormatio n Technolo gy Mazandara n U ni ve rsity of Science an d Technolo gy Babo l, Iran areihanian@us tmb .ac.ir Beh rouz Minaei-Bidgoli Departme nt of Computer Sc ience Iran U niversity of Science and Technolo gy Tehran, Ira n b_mi na ei @iust.ac .ir Muhamm ad Yousefnezh ad Departme nt of Computer Sc ience Nanjing U niversity of Aeronautics a nd Astro na utics Nanjing, China myo us ef n ezh ad@ nuaa.edu .cn Abstrac t —Find i ng me aningfu l com munities in socia l netwo rk has attracted the at tention s o f man y resea rchers . T he comm unit y structu re o f co mplex n et wo r k s rev eals bo t h their o rganiz ation and hidden rela tions amo ng their cons t itue nts. Most of the researc hes in t he field of comm unity d etectio n ma i nly focus on the topo l ogica l st ructur e o f the ne t wo r k wit ho ut p erformi ng any content ana lysis. No wadays, real wo rld social n etwork s are contain ing a v ast range of i n form atio n including shar ed objec t s, com ments, f ol l ow ing info rmat ion, etc. In re cen t y ears, a num ber of research es hav e propo sed appro aches wh ich consider bo t h the content s that are in t erchang ed in the netwo rks and t he topo l o gical structur es of t he network s i n order to find mo re meanin gful co mm unities. In this res earch, the effe c t of t op ic analysis in fi nd ing mo re meaning f ul com munit ies i n s o cial netwo rking sites in wh ich the us ers express t heir feeling s towa rd different ob jects (like mo vies) by the means of rating is demo nst rated by pe rform ing extens ive expe ri men t s. Keywor ds-Con te n t An alysis; Topi cal co m muni ty; Co mmunity detection ; Modular ity; Pur ity I. I NTRO DUCTION With the emergence of social n etw orks, peo ple h ave bee n attracted to t h em , and hav e been sharing valuab le info rmatio n by means of c ommunicati n g wit h e a c h o th e r. F or exampl e, folkso n omies ar e so cial tagging sites which their users co l laborative l y e xpress their f eeli n gs an d se ntiments tow ar d a spec i al r eso ur ce like a mov ie or m us ic by m eans of desc r iptiv e keywo r ds (tags) [1] o r rati n gs. O ne of th e most impo r tant issues c o n side r ed when analy zi n g th ese kinds of netwo r ks i s co m munity dete ctio n . A Community (al so so m etimes r ef erred to as a module or cl uster [2]) is a dense s ub-netw ork w ithin a larger netw ork, such as a clo se-knit group o f f r i ends i n a so cial netwo r k [3]. The commu nity str uctu r e of complex netw o r ks reve als bo t h thei r o r ganizat ion a nd hi dde n relat ions a m o n g their co nstituents [4] . A large n umber o f m et h ods ha ve b een p r opo sed t o extract appropriate co mm unit ies from netwo r ks including a set of nodes (indiv iduals) and w eighted edge s (connectio ns). They just co nsider the g r aph struc ture of t he n e t w or k fo r fin di n g co m muniti es and n o c onte nt a na ly sis has bee n used in t h e proce ss of th eir p ropos ed approaches. Des pite of the orig inal defi n itio n of the netwo r ks, now aday s, r eal w orld netwo rks like F a ce boo k and T w itter are co n taini ng a v ast range of info rmatio n in cl udi n g s hare d obj ects, co mme n ts, etc . It is unre asonab le f or a com munity to be expl ained by a s ingle e ntity b ec ause t h e co mmuni t y membe rs are ge n e rally inte racti n g wi th eac h ot h e r via a l arge numb er of dist inguis hable way s in v ario us do mains. O ne of the pos sible solu tio n s is to find topic al clus ters in w hi c h the n o des have t h e s ame t o pic of inte r es t. Eac h topic al cluste r rep r ese nts o n e of th e topics of inte r est in the ne t w o rk. The n, a comm unity detec tion algo rit hm c an be app lied to these topica l cl uste r s to find th e ulti mate co mmu nities [5]. In th is w a y , we ca n analy ze and es t im ate th e eff ect of t opic co n side rat ion in co mmuni ty dete ctio n. In this pape r , th e ef fec t of topic analy sis in find ing m ore meani ngful co mmunitie s in soc i al netw orki ng site s i n w hic h the use r s express the ir fee l ings tow ar d diff erent obj ects (like mov i es ) by t he m ea ns o f rati n g, i s de m ons tra ted by per fo rming exte n sive expe rime nts. The r efo r e , a n e t w ork is p art i tio n ed into diff er e n t topic al cl usters in w hich the n odes hav e t h e s ame topic o f i nte res t. Th e n, a co mmuni t y detec tio n a lgo rithm i s appl ied to the top ical clus te rs in o rder to f ind m o r e mea ningf ul co m muni t ies . This wil l l e ad us to com m uni ties in whic h th e nodes ar e t ight ly connecte d and hav e t he sa me top i c o f inte r e st. This proc ess is c alled top ic-o rie nt ed comm u nity dete ct i o n [5]. At last , the resu lt s of com muni t y de tect i on wi th top ic co n side rat ion are co mpa r ed w ith th e r esu lts of co m muni t y detec t io n w ithou t c o n side r ing the topics of inte r e st. Quant i ta t ive ev a lua tio ns rev eal that the resul ts o f co mmu nity detec t io n wi l l be im prov ed when the topic of inte r est in th e netw ork is co nside r ed . The r ema inde r of the pape r i s out lined a s fo ll o w s. I n sec t ion II , r e lated w orks are r ev iew ed. S ec tio n III expl a ins the topic-o riente d co m mu ni ty detec tio n. In o r der to ev alu ate the eff ect of topic co n s i de ra tio n in ide ntify in g th e c om m uni ti es of rating-b ase d soci a l ne t w orks, exte n s ive expe rime nts are co n duc ted on r eal -lif e da t a se ts. The de sc riptio n s of t h ese d ata sets, th e expe rimenta l r es ults and thei r analy ses are gi ve n in sec t ion IV . Fi nally , the co n c l us ions ar e giv en in se ction V . II. R ELATED WOR KS Many r esearc hes have been do n e in th e area of c ommunity detec t ion. Most of th ese r esearc h es mainly focus o n th e topo l o gical structure o r linkage p atterns o f netwo r ks. They merely co nsider the graph st ructure o f the n etwo r k fo r findi ng co m muniti es , while n o co n tent analy sis i s used in t h e proc ess of their p r opo sed approaches. Acco r ding t o the commu nity dete ction strategies whic h were employ ed in these r es ea r ches, thei r p r opo sed m et hods can b e classifie d into optimizatio n -based methods and heuristic met hods. So me of the optimizatio n -based methods foc us on o pt imi zing an o bjectiv e function [5]. One of the most important w orks in the l ite ratu re w a s a r ese a rch done b y New man a nd Gi r v an, in w hi ch they introduc ed mo dularity as an obj e ctive function [6]. A la r ge amou n t of works has b een done to optimize modularity such as the methods w h ich we r e deve lope d in [7-9]. This function has b ee n influe ntial in the literatu r e o f community dete ctio n , and has gained succ ess in many applic ations. Modularity is use d to evaluate the qu ality of a pa rticular divisio n of a n etw ork into co mmunities [5]. On the ot her h and, heu ristic met hods such as GN algorit hm [10] and CPM algo r it hm [11] des ign a grap h cluste r ing algo r ithm base d on intuitive assumptio n s [5]. Eve n t h o ugh these r ese arc hes have g aine d suc ce ss in some appl ica tio ns, si n ce they mainly focus o n the to pologic al struc t u r e of the netw orks, they igno r ed the co nten ts inte r c han ged b etw een membe r s. As a res ult , the r e lat io n shi ps be t w e en the memb ers i n t h es e r ese arches are mainly base d on the tota l nu m be r of commu nicatio n s. In recent y ear s, a num be r of resea r ch es h ave propos ed approaches w hich consider bo t h th e co nt e n ts that are inte r cha nged in the n etw orks an d the topo logical structures of the netw o r ks in orde r to find more meaningf ul commu n ities . Z. Z hao, e t a l. p ropos ed a topic-o r i e n ted co mm unity dete ction approach b ased on soc ial objec ts’ clustering and li n k analy s is [5]. Their propos ed approach could ide n tify th e topical co m muniti es whic h r efle ct t he to pics and st r engt h s of co n n ectio ns si multaneously . Zhu, et al. c omb in e cl a ssic i de a s in topic mode l ing w i th a v ar ia nt o f m ixed-memb ers h ip block mode l w hi c h is rece ntly de ve l o p ed in the statistic al phy sics co m munity [12]. In th ei r resea r ch, Z hu, et al. co mbine topic - mode ling wit h link st ructu r e. A. Z ha o a n d Ma propos ed a framew ork to appl y a semantic ally structured app r oach to th e Web service community modeling and disc ov er y [13]. III. T OPIC - ORIENT ED COM M UNITY DETECT ION IN A SOCIAL NETWOR K As w e m e n tio n ed earl i e r, th e go al of this pape r is to demo n st rate the ef fec t of topic cons ide ra tio n in fi ndi n g more meani ngful co mmunitie s in soc i al netw orki ng site s i n w hic h the users ex press thei r fee l ings tow ard diff er e nt ob je cts by t he means o f rating. For this c ause , so me compo nents of t he f rame w o r k w hic h was p ropos ed in [5] ar e cha nged in o r de r to be appl icab le to the me nt io n ed soc ial n etw orks. This f r amew ork detec t s com muni ties w h i c h have u nique top i c of inte r es t and co nn ected m e mbers. Eac h co mmunity co n tains the node s of the netw ork w hic h have t he same topic of inte res t. This f r amew ork is imp leme nted i n fo ur steps : P r ep r oce ssing and annota ting topic labe l s, Clus t e r ing so cial ob jec ts, Cre a ti ng topic a l clus te rs and A pply ing a com m uni t y dete ctio n algo rithm t o th e topic a l cluste r s . A. Preprocessing and annota t ing topi c lab els In this step, data se ts are p reproc ess ed and r eady to use. In this p r oce ss, the so cial ob jects are r eco gn ized. Gene rally , Peo ple communicate w ith eac h other t hr oug h soc i al o bjects. These objec ts often imply the topics which peo ple ar e inte r este d in. So cial ob jects can be classif ied into tw o kinds of situatio n s [5]: 1) the soc i al ob jects which are a ttac h ed to multi-memb er s, 2) the so cial ob jects which are attached to one membe r . In the fi r st situatio n, the edges b etwe en m emb ers are b ui lt bec ause of a soc ial o bjec t. A n example of this situatio n can be happe ned in a mov ie r ating n e two r k. In this netwo r k, edg es be t ween members are built w h en they r ate the s ame mo vie . As a matte r of fact, in thi s n etw ork, eac h mo vie (soc ial obj ect) is attac hed t o multi m emb er s . The membe r s of the mov ie r a ti n g netwo r k are co nn ec ted to each o the r due to th e r a ti n g of th e same mo vie. In the se co n d situatio n, each soc i al o bject i s attac h ed to only o n e member. Theref ore th e so cial obj e cts ar e c onsidere d to be th e attributes of membe rs of th e n etw ork. An examp le of this situation can b e ha ppe n ed in a p ape r cita tio n n etw ork. In this n etw ork, p a pers (me mbe r s) cite eac h o th e r. Also , e ach paper c ontains a text co nt ent ( th e ti tle o f a pape r ) whic h is a social o bject and c an be co n sidered as the a tt r i b ute o f the co r r espo n ding pape r . Figure 1 show s th e tw o different ki nds of r elatio n s be t ween the memb ers of a n etw ork and soc i al ob j ec ts. Th e netwo r k w h i c h i s loc ated in the left side o f Figure 1 is a movie rating n etw ork. As it is clear, the edge s be t ween membe r s are built b ecause of t he so cial ob jects. Also , the n etwo r k w hi ch i s loc ated o n the right side of t he Figu re 1 is a p ape r cit atio n netwo r k. In this netw ork, each so cial obj e ct is the a ttrib ute of its correspo n di ng paper. Since in this paper th e social n etw orking sites in w hi c h th e users exp r ess t h ei r fe elin gs t ow ar d diff er ent ob j ec ts are analy zed, the fi r st situatio n is happe n ed. Fig. 1. Two di ff eren t kinds of r elatio ns be t wee n th e me mbe r s of netw orks and so cial o bjects So , in t h is step , data sets ar e preproc essed an d r eady to use. I n this proc ess, the so cial ob jects are r eco gn ized. Af terward, the topics o f e ach soc i al o bject in the da t a s et are retriev ed. Sub sequently , e ach soc ial ob ject is labele d b y its co r r espo n ding t o pic. In so me cases th e topics of each so ci al obje ct can e a sily be r et riev ed manually , or there are co r r espo n ding tags which represe nt the topics f o r eac h soc ial obje ct. But in case s w h ere a so cial ob ject is r epresente d b y text and its la be l s cannot easily b e r etrieve d, a method ha s bee n int r oduce d by Z. Zhao, e t al . w hich c an a nn otate the topic labe l to eac h social o bject [5]. B. Clustering soc ial obje cts In this step, so cial ob j ec ts in a n e two rk a r e parti tio ned into diffe rent cluste r s . Eac h cl uster represe nts a unique topic which is shared by its memb ers. In ot h e r wo r ds, Acc ording to their labele d t o pics, social ob jects ar e pa r titio n ed into diff er ent clusters in a way t hat e ac h c luste r includes m emb ers with th e same topic. Diff erent m et h o ds can be used to pe r fo rm the soc ial ob je ct cluste ring acco r ding to the ty pe o f soc i al obj ects. Fo r e xa m pl e, a n ov el m et hod has be en propos ed in [5] to cluste r the te xt so cial ob jects. This met hod comb in es the ve cto r spac e m od el w ith the Ent r opy Weight ing K-Me ans ( EWKM ) [14] in o r de r to clus t e r the text so cial ob je cts. Since the d ata se ts whic h are used in this pape r cont ain so cial obj ects with lab eled top ics, we ma nually parti tio n t hese so cial o bj ects into dif f ere nt clus ters. C. Creating top i cal clus t ers U sing the resu l ts t hat are ge n e rat ed in the prev i ous ste p , we part itio n the me mbers of th e ne t w ork into diff er e nt t opic al cluste r s . In the f irst step , each soc ial obj ect h as b ee n anno t at ed w ith a t o pic lab el. In this ste p, me mbe rs are pa rt it i oned into diff er e n t to pic a l cl usters w ith co n side ring t h e top ic l a be l s of the so cial obj ects they ar e invo lved i n. Thus in this step we fi nd cluste r s in w hic h ev e r y m e m be r has the same topic o f inte r e st. The refo r e the t ota l numbe r o f t opic al c luste r s is eq ual t o the numb er of topics of inte r est in th e n e t w o r k. A use r ca n be a membe r o f sev er al top i ca l cluste r s , si n ce it is co mmo n fo r a user t o b e inte r este d i n se ve ra l topics . D. Applying a communi ty detect ion algor ithm to the topical clusters This step aims to f ind co m munities in each of the topical clusters w h ich we r e created in the prev ious step. Me m be r s in each topic al clus ter ar e co nn ected to each ot h er w i th diff er ent strengt hs. B ased o n the number of r atings on the same so cial obje cts, some members may have st r o n ger co nn ections , w hile so me o th er may have w eak o r no c onnec tions. This h as been co n cluded acco r din g to the t opic analy sis th at has been perfo r med in the f r ame wo r k. Si nce the result o f th e framew or k is to detect co m munities w hi c h have unique topic o f interes t and c onnected membe r s, we should apply a co mmuni t y detec t ion algorithm to th e prev i o usl y cr eated topical clus ters in order to identify th e tig h tly connec ted membe rs. In o r de r to per f orm this proce ss, man y c ommunity detec t ion algorithms ca n b e employe d suc h as GN and so on. New man pr o pos es an impo rtant algorit hm t o pa rtiti o n n etwork graphs o f links an d n ode s i nt o sub g raphs. H e also int roduce s a concept w hich is called modula r ity . In t h e case of we igh te d netwo r ks, modularity ha s be en def ined as f ollow s [9]:              j i j i j i ij c c m k k A m Q , , 2 2 1  (1) Where A ij r ep r esents the w eigh t of th e edge be twee n i and j,   j ij i A k is th e sum of t he we i ght s of t he edges attached to vertex i, c i is the co mm unity to which the verte x i is assigned, the δ function δ (u, v) is 1 if u=v and 0 o the r wise and also   ij ij A m 2 1 . Since New man 's algorithm was ve r y ti me-co n suming, Blo n del, et al . sugges t the modif i ed v ersion of the al go r it hm in order t o make it faster, giving r ise to w h at i s know n a s the "Lo u vain m et h od" [15]. This algorithm is a m odula rity maximizat ion algo rithm which iterat ive l y optimizes the modularity i n a local way an d aggregates n ode s of th e same co m munity [16]. In this pape r , the "Lo uv ain method" has b een applied in o rder t o f ind topic al co m munities. IV. E XPERIM ENT AND ANA LYSIS In this sec tio n, the r esu lts of our resea r ch are prese nted. Firs t, three rea l-life d ata sets along wit h a perfo rmance m et ric w hi c h w er e use d in th e expe rime nts ar e des c ri be d. The n, the proc ess o f detec ting topic al comm u nities i n the m e nti o n ed da t a sets is disc ussed a nd its resu l ts a r e analy zed. Fina lly , th e resu lts of topic-o ri e nted co mmu nity detec ti o n (w it h perfo rm ing co n tent ana ly sis) are comp ar ed w it h the resu lts of commu nity detec t io n w i tho ut perfo rming any co nt e nt analy sis. A. Re al life data se ts We used th e public ly availab le dat a sets in ou r e xpe ri me nts w hi c h are Mo vie le n s, Boo k-Cr oss ing and CIAO dat ase t s. Mo vie lens data se t [17] is a rating data se t w hi c h is c ollect ed from the Mo v iele n s web s ite ( http:/ /mov i elens .o r g). I t co n sis ts of 100000 ratings f r om 943 use r s whic h w ere give n t o 1682 mov i es. B ook -Cr os si ng d ata se t [18] is a rati ng d ata set w hich is col l ec t ed fro m the B o ok-C r oss ing co mmuni ty (http ://w ww . boo kcrossing . co m). It co ntai ns 278858 use r s prov iding 114 9780 ratings a b out 2713 79 b oo ks. CIAO da ta s et [19-22 ] is a ra ti ng dat a se t w h i c h is co llecte d f rom a prod uct review site ( http ://ci a o.c om ) in whic h use r s share the ir o pinio n s abo ut a product by mea ns o f r ati n g o r co mme n ting . The r e are 35773 rat ings in this data se t w hic h ar e attac h ed to 168 50 produc t s by 2248 use r s . As desc r i b ed earlier in th is pape r, the t opic -o r ie n t ed co m munity dete ctio n f ra mew ork considers the r esults of topic analy sis fo r fi n ding mo re meani n gful co m munitie s. So , in order to ev a luate thi s f r amew ork, two aspec ts should be co n sidered: topic and lin kage structure. It means that t he expe cted r esults should kee p eac h co mmunity 's membe rs wi th the same topic and strong co nnections. Z h ao et al. introduc ed a pe rformance evaluatio n metric w hich co n si ders bo th topic and linkage st r uctu re [5]. This metric h as be en def in ed as follow s:        Q Pur i ty Q Purity Pur Q      2 2 / 1      As it is clea r in the abov e equati o n , The PurQ β has three parameters w hich are Q, Purity and β. Q de notes t he modularity . T his para meter measures the commu nities f r o m the perspective of the link s tructure . T h e larger th e Q, t h e be t ter t h e co m munitie s are div i ded from t h e perspec t ive of topo l o gical str ucture . In our expe r ime n t, fo r each topical cluster, modula r ity is calculat ed by equation 1 . S ince th e topic-o r iented framewo rk may gener ate more than one topic al cluster f or eac h data se t, t he total value o f modularit y in this framew ork is calcul a ted as f ollow s:      n i TC T TC i i Q W eight W eight Q 1   Where n is t h e numbe r of generated topica l cluste rs. i TC Q is the v al ue of modularity fo r the topical c luster TC i . i TC Weight is the su m of t he weights of edge s in the to pical cluster TC i . We ight T is the sum of the w eights of edges in the topical cluste r, w hich is directly created from the basic netwo r k (when n o topica l cluste r ing has b ee n pe r fo r med). It should be co nsider ed that since in this f r amewo r k n o co m municatio ns' co n tent analy sis is perfo rmed, t he w eig ht of each r elatio n ship be twee n t w o membe r s is the n umb er of ratings w hi c h are giv en t o t he same so cial ob jects by th ese two membe r s. In equa tio n 2, Pu rit y r ep r esents t h e pu rity o f to pics in the detec t ed commu nities a n d is calculate d as fo ll ow s [5]:        cm N i i ij k j cm n n N Purity 1 1 } / { m ax / 1   Where N cm repr ese nt s the num be r of dete cte d co m muniti es , n ij r ef ers to th e n umb er of n ode s b elonging to topic j and co mm unity i , n i refe r s to the numb er of nodes in co m munity i. k is the numbe r o f topics in the n etw ork. Th e highe r the Purity , the better th e commu nities are partitio ned from the pe r spectiv e of topics. β is a paramete r to adjus t the w eight of Pu rity a nd Q and β ] , 0 [   . If w e co nsider the puri t y of to pics and the to pology of the n e tw ork equa lly i mpo r tant , the v alue of β s h ou ld be set t o 1. If we want t o pay mo r e atte n tio n to Puri ty in co mparis on w ith Q , th e n th e v alue of β s h ould be se t to a numbe r be t w ee n 1 and ∞. On the ot h er hand, if we w an t to pay mo r e atte n tio n to Q in comp ariso n w ith Pu rity , th e v alue of β s h ou ld be set to a numb er betw een 0 and 1. Ac tually β is use d in eq uatio n 2 to adjus t the e mphas is of topics and li nk struc tu r e [5 ]. B. Ex perime nt s In orde r to ide n tify t h e commu nities by appl y in g the topic- oriented co mmunity detec t ion f r amewo r k to t he thre e introduce d datasets, fo ur s teps (acc ording to section III) h av e bee n take n. T h e f irst s tep w as to preproc ess the d ata sets. As to the Mov i elens an d B oo k-Cr ossing d a ta sets , mov i es an d boo k s were conside r ed as the soc ial o bje cts. So , for t h e Mov ielens data set th e genres of th e mov ies were extracted. These extracted ge nres ar e the s ame as the genres attac h ed to each mov ie by IMDB (h tt p:/ / w ww . imdb.co m ). Then, all t h e mov i es w h ich were in th e ge nr es of Do cumentar y o r W estern were r etrie ve d. As yo u kn o w , th e genr e of a mo vie repr ese nt s the gene r al t o pic in which a mov i e is m ade abo ut . In this step, we a chie ve d 77 movie s. For th e B oo k-Cr oss in g data set, we extracted the c ategories of 93 boo ks from A mazo n (http://ww w .amazon.co m). As fo r the CIAO data set, produc ts were considered as the soc ial o bjects. Each produc t's category w as at tached to it in the data set. Thus fo r t h e Bo ok-Cros sin g data set and the CIAO data se t, the categorie s r ep r esent th e topics of eac h produc t o r bo ok. The sec ond step w a s to cluste r th e social ob jects. As f or the Mov ielens data se t, the m ov ies we r e partitioned into two clusters of Docume n tary a nd Weste rn. The Do cumenta ry cluster contained 50 mov ies whil e the Wes tern o n e containe d 27 mov i es. As fo r the Boo k-Crossing data set, the boo ks were partitio ned into two clusters of Fiction and No n -Fictio n . Th e Fictio n cluste r co n tained 80 boo ks, w hi le th e Non-Ficti on cluster conta ined 13. T he p r oduc ts i n t h e CIAO da ta set we r e partitio ned into six cluste r s of DVDs, B ooks, B eaut y , Music, Travel, and Foo d an d Drink. Th e DVDs cluste r co n tains 2057 products, The Bo oks cluster contai n s 2803 products, th e Be a ut y cluster contains 2333 produc ts, the Music clust er co n ta ins 1801 produc ts, the Trave l cluster co ntains 3922 products an d fi nally th e Foo d and D rink cluste r c ontains 3937 products. The t hird step w as to c r eate topical cl usters. Theref ore in each data set, the use rs who ra te th e soc i al objec ts in ea ch cluster we re partitioned into topic al c luste rs. For e xample, all users who r ate th e movie s in th e cluster of “ D oc umentary ” were p ar titio n ed into the t o pical c luste r o f “ Documentary”. The me mbe rs o f each to pical cluste r rated the soc ial ob jects w hi ch have the same topics . Thus acco r ding to the n umb er of topics , we achi ev ed two t opic al cluster s f o r the Mo vie lens and Boo k-Cr oss in g data sets and 6 topic al clus t ers fo r the CIAO data set . As m e nt ioned earlier, si n ce i n this f ram ew ork no co m municatio ns' co n tent analy sis is perfo rmed, t he w eig h t of each r elatio n ship be twee n t w o membe r s is the n umb er of ratings w h ich ar e given to the same soc ial ob jects (for example, t w o m ovie s in the genre of Do cumentary ) by th es e two m embers. The las t step was to dete ct to pical co mm unit ies. T hus w e applied the "Lo uv ain method" t o eac h topical cluste r create d in t he previo us step. In orde r to acc ura tely calculat e t he modularity , we a pplied th e Lo uvai n m et h od t o each topical cluster t e n t imes, and ca lculate d th e ave ra ge o f th e achieve d values of modula rity . Table I giv es th e r esults achieve d by appl y in g th e topic- oriented co mmunity de tection f ramew ork to the Mov ielens, Boo k-Cr oss in g and CIAO d a ta sets . I n this Tab l e, the co lumns "Topic al Cluste r s", " No. of Edges" and "No. of Nodes" represe nt th e c r eated topica l cluste rs in the p r o ce ss of appl y in g the t opic -orie nted framew o r k to the three mentio n ed data sets , the numbe r of edge s and th e numb er of n odes e xisting in eac h of these to pi cal cl uster s , respe ctively . Mo reove r, the co lumn s "Total Modula rity " a nd "Pu rit y " denote the o verall modularity value (Q) and Pu r ity value fo r all o f th e to pical communit ies. TABLE I. T HE RESULT S ACH IEVED BY APPLYING THE TOPIC - OR IENTED COMMUNITY DETECT ION FRA MEW ORK TO M OV I ELENS , B OOK -C ROSSING AND CIAO DATA SETS Purity Tota l Mod ularity No. o f Nodes No. o f Edg es Top ical Clusters Data sets 1 0.12 44 352 158 33 Document ary Mov ielens 491 693 69 Western 1 0.84 69 102 1 8531 Fiction Book - Crossi ng 191 158 7 Non-Fiction 1 0.30 86 135 6 539 16 DVDs CIAO 904 899 9 Books 811 526 7 Beauty 569 207 6 Music 867 129 05 Travel 119 3 297 63 Food & Drink As it is clear in Tab l e I, Puri ty has i ts maximu m v a lue in eac h of the thr ee data se ts. The r easo n is that , th e topic al cluste r s crea ted in e a c h dat a se t inco rporate m e m be r s w hi c h are inte r este d in the s ame unique t o pic s. The r efo r e the puri ty of topics in e ac h of t h e to pica l c om m uni ties is 1 acco rding to equa tio n 4. It s h o uld be c o n si de r ed tha t it is po ssible f o r a ce rtain use r to b e i n sev eral topic a l c l us te rs, s in ce the inte r e st of pe ople in sev eral diff erent t opic s is com m o n . Thus some of the membe rs of topic al clus t e r s in eac h data se t m ay be th e same . F o r ex a mple , co n side r th e c ase tha t a us e r ra ted se ve ra l diff er e n t mov ies. So m e of these mov ies w ere in the genre of Do cument ar y , and the o thers w ere in the ge nre of W este rn. The refo r e t his use r b elo ngs to bo th t op ical cluste r s in the Mo vie lens d ata set. C. Comparison In o rder to prov e th e supe r io rit y of t h e resul ts o f de tecti ng co m muni t ies with top ic co n side ration, in this sect i o n, we co m pare the r es ults o f topic-o rie n ted co mmuni ty detec tion, w hi c h was imple me nted in sec tion B , w ith t h e r es ults of Classic al Com muni t y Detec tio n in w hic h n o co nte n t ana l y sis is perfo rm ed . In the proce ss of cl assic a l com m uni t y detectio n appro ac h, a co m muni t y detec tion a l go rithm is applie d to a netw ork (b a s ic netw ork) in w hic h the we i g ht o f th e edge s represe nts the numb er of co mmunic atio n s be t w een relev ant nodes . In th is co n ditio n, no c onte n t ana ly sis i s do n e . We first applied the "Lo uvain method" to the basic netwo r ks of th e M ov ielens, B oo k-Cr oss in g a nd CIAO data sets (impleme nting the Classical Commu n ity Detec tion Framew ork). Then w e partitio ned the basic netw orks of t he three me ntioned data sets into t opic al cluste rs. Each topical cluster in cludes memb ers w hich have th e same t opi c. Af terwards, t he Lo uvain m et hod w a s applied to t hese to pi cal clusters (i mpleme n ting the Topic-o r iente d commu nit y detec t ion framew ork w hich w as discusse d in section B ). We then use d Pu rQ β to e v aluate the pe r fo r mances i n the expe rimental evaluatio n. The co rr espo n di n g r esul ts a r e give n in Table II. Conseque ntly , as i t i s show n in Table II, β was set to 0.5, 0.75, 1 , 1.5, 2 r espec t ive l y , w h i c h represe n ts th e diffe rent stre n g ths fo r the t opic and the link. Pu r ity , Q and PurQ β ha ve bee n ca l culate d f or e ac h o f th e tw o mentioned framew orks. Acco r ding to Table II, Modul ar ity an d Purity has hi ghe r values in th e t o pic-o r ie n ted f ra mew ork, since th e basic netwo r k i s partitio ned int o topical cluste rs, and each identifie d co m munity in clu des me mbe r s who have the sa me topic of inte r est. Th e r efo r e, the t opic -orie nt ed co m munity dete ctio n framew ork has a highe r value of Pur Q β fo r all five values of β. T ABLE II. C OMP ARISON OF M ODULARIT I ES W HICH WERE ACH IEVED BY APPLYING THE TOPIC - OR IENT E D FRAMEWORK ALONG WITH CLASSICA L COMMUNITY DETECTION FRAMEWORK T O EACH O F THE THREE MENTION E D DATA SETS . PurQ β Tota l Purity Tota l Mod ularity Framew orks Data set β=2 β=1. 5 β=1 β=0. 75 β=0. 5 0.13 21 0.14 95 0.1955 0.2519 0.3 760 0.97 77 0.10 86 Classical Mov ielens 0.15 09 0.17 03 0.2213 0.2830 0.4 154 1 0.12 44 Topic-ori ented 0.85 02 0.85 72 0.8699 0.8795 0.8 906 0.90 50 0.83 75 Classical Boo k-Cro ssing 0.87 37 0.88 88 0.9171 0.9389 0.9 651 1 0.84 69 Topic-ori ented 0.33 32 0.36 24 0.4294 0.4963 0.6 038 0.82 79 0.28 99 Classical CIAO 0.35 81 0.39 20 0.4716 0.5535 0.6 906 1 0.30 86 Topic-ori ented V. CONCL USION This pape r ev alua tes the ef fect of topic co n sideration i n finding mo r e meaningful commu nities i n soc ial netw orkin g sites in w h ich the use rs express their fe elings tow ar d dif fe r ent obje cts (li ke mov ies) by th e means of rating. Th e r efo r e, the netwo r k is p artitio ned into dif f erent t opic al cluste r s in w hich the n odes have the same topic of inte r est. Then, a co mm uni t y detec t ion algorithm is applie d to the topical cluste r s in order to detec t commu n ities . After tha t, a co m parison ha s bee n perfo r med b etwee n th e results o f topic -orie nt ed commu nity detec t ion and th e r esults of Classical Commu nity Detec tion in w hi ch n o co nt e n t an aly sis is perfo r med. The experime ntal results indicate that th e r esul ts of topic-o r iented co m muni t y detec t ion will be i mprov ed w h en it is jo in ed wit h topic analy sis. The re is a plenty o f roo m to s tudy on co mmuni t y detec t io n prob lem in r ea l comple x n e two r ks w hic h co n ta in h uge am o unt of info rmatio n wi th diff erent natu res. The refo r e, in fu tur e w o r ks w e hav e a plan to w ork o n th e eff ect o f o th e r ki nds of co n tents in the netwo rk, like the co mmu nica tio ns' co nt e nt analy sis, in f ind ing mo r e me ani ngf ul c om m uni ti es in so cial netw orking s ites in w hich the use r s ex press th e ir fe eli ngs tow ard diff ere nt ob jects w ith rat ing. R EFERENCE S [1] Chakra borty, A ., Gho sh, S., & Ganguly, N ., 2012 . Detecti ng overla pping communitie s in folkso nomies, Pro ceeding s o f th e 23 rd ACM co nferenc e on Hypert ext and social media . Publis hing, pp. 21 3-218 . [2] Leskov ec, J., Lang, K.J., & Mahoney, M., 201 0. E mpirical c ompari son of al gorith ms for network co mmunity d etection, Pr oceedi ngs o f t he 1 9th international con ference on W orld wide web. Pu blishing, pp. 631 -640. [3] Ne wman, M . 2011 . Co mmunitie s, m odules a nd large -scale structur e i n networks. Na ture Physics, 8 , 25-31. [4] Lan cichinetti, A., & Fortu nat o, S. 2 012 . Conse nsus clu stering i n complex networks. Sc ientific rep orts, 2 . [5] Zhao, Z ., F eng, S., Wan g, Q ., Hu an g, J.Z. , Willia ms, G .J., & Fan, J. 201 2. T op ic orient ed commu nity d etection t hro u gh s ocia l objects an d link a nalysi s i n so cial networks. Kno w ledg e-Base d Systems, 26 , 164- 173 . [6] Ne wman, M.E. , & G irva n, M. 2 00 4. Finding a nd evaluating commu nity structur e in netw orks. Physical re vie w E, 69 , 026 11 3. [7] Are nas, A., Du ch, J., Fer nández, A., & Gó mez, S. 200 7. S ize reduct ion of c ompl ex net works preservin g m odularit y. New Jo urn al o f Physics, 9 , 176 . [8] Leicht, E.A., & Ne wman, M .E. 200 8. Co mmunity structure in dire cted networks. Physical rev ie w lette rs, 100 , 118 703 . [9] Ne wman, M .E. 2 004 . An al ysis of we ighted netw orks. P hysic al Review E, 70 , 056 13 1. [10] G irvan, M., & Ne wman, M.E. 2 002 . C ommu nity structure in social a nd biological n etwork s. Proc eed ings o f the Natio nal Aca demy of S cienc es, 99 , 78 21 -7826. [11] Palla, G., D eré nyi, I., Farkas, I., & V icsek, T . 20 05. Uncovering th e overlappi ng community structure o f complex network s in na ture an d society. Nature , 435 , 814-818 . [12] Zhu, Y., Y an, X., G eto or, L., & Moor e, C. 20 13. Sc a la ble texta nd l ink analysis with m ixed -topic li nk mode ls. Pro ceeding s of th e AC M SIGKDD I nternational C onferen ce on Kno wledge Disc overy and Data Mining ( KDD). Pub lishing, pp. 473 –4 81 . [13] Zhao , A., & Ma, Y., 2 012 . A Sema ntically S tructu red Appr oach to Service C ommunity Di scovery, Semanti cs, Kno wledge and Gri ds (SKG), 201 2 Eighth Inter national Co nferenc e on. Publi shing, pp. 136 - 142 . [14] L. Ji ng, M . Ng , J. Hu ang, An entrop y weighting k-mea ns a lgor ithm for sub space clus tering o f h igh-dimen sional spa rse da ta , IEEE T rans a ction s on Knowle dge and Data Engi neering, vol. 19, pp. 102 6–1 041 (20 07) [15] Blo ndel, V.D., Guil laum e, J .-L., La mbiotte, R., & Lefebvr e, E. 20 08 . Fast unfol ding of c ommuni ties in large n etwork s. Jo urn al of S tatistical Mecha nics: The ory a nd Experime nt, 200 8 , P100 08. [16] Wa ng, D ., Kwon, K., So hn, J ., J oo, B. - G., & Chung, I .-J., 201 4. Community To pical “ Fingerpri nt” Analysis Ba sed on S ocial Sema ntic Network s, Ad vanced T echnol ogie s, Embedded a nd Multim edia fo r Human-c entric Co mputing. Pu blishi ng, pp. 83 -91. [17] H erlocker, J.L ., Konstan, J. A., B orc hers, A., & R i edl, J. , 19 99. A n algorith mic framework for per formin g collab orativ e filt ering, Proceeding s of t he 22n d annual internatio nal ACM SIGIR conference on Research an d develo pment in in formatio n r etri eval. Publishi ng, pp. 2 30 - 237 . [18] Zie gler, C. -N., McNee , S.M., K onst a n, J .A., & Lau sen, G. , 20 05 . Improvin g r ecommendati on li st s throu gh topic diversi ficatio n, Proceeding s of the 14 t h internat ional conferenc e o n World Wide W eb. Publishin g, pp. 22 -32. [19] T ang, J., Gao , H., Hu , X. , & Liu, H ., 201 3. Exploi ting homophily e ffect for tru s t pr ediction, Pro ceeding s o f t he sixth ACM internati onal confere nce on Web s earch and data mini ng. Publ ishing, p p. 53 -62. [20] T ang, J ., G ao, H. , & Liu, H., 20 12 . mTrust: d iscernin g multi -facete d trust in a connect ed world, Proceedi ngs of t h e f i fth ACM inte r national confere nce on Web s earch an d data mini ng. Publ ishing, p p. 93 -102. [21] T ang, J ., Gao , H., Liu, H ., & Da s Sar ma, A., 20 12. eTru st: Understan ding t rust evolution in an online world, Proceedi ngs o f th e 18th A C M SIGK DD internat ional conference on Knowledge discovery and data mini ng. Pub lishing, p p. 25 3-261. [22] T ang, J., H u, X . , Gao, H., & Liu , H., 20 13. Exploiting l ocal and global social context for r ecom mendatio n, Proceedi ngs of th e T we nty -Third international j oint confere nce on A rt i ficial I ntelligen ce. Publishin g, pp. 271 2-2718.